Indoor occupancy estimation using environmental parameters
Khalid Masood, Mustafa
Date of Issue2017-08-23
School of Electrical and Electronic Engineering
In this work, we address the problem of obtaining reliable estimates of the real-time occupancy for Air-Conditioning and Mechanical Ventilation (ACMV) systems. We take a non-intrusive approach that is gaining popularity, which is the use of environmental parameters such as CO2, temperature, humidity and pressure. The approach requires the extraction and selection of useful features from the sensor data, which can be used with machine learning techniques to obtain occupancy estimates. To date, occupancy estimation has primarily focused on static features, which we define to be features sampled instantaneously, as opposed to dynamic features that are sampled over a moving window. In this work, we propose novel feature selection techniques that employ both static and dynamic features. We study the performance of these algorithms with experimental data collected through a multi-sensory network in an office space. First, we propose novel feature selection algorithms within the static feature paradigm. In past works, feature selection has generally been implemented using filter-based approaches. In this work, we introduce the use of wrapper and hybrid feature selection for occupancy estimation. Compared to filter methods, our approach can achieve a better occupancy estimation accuracy. Additionally, we use a ranking-based combinatorial search in our algorithms, which is more efficient than the exhaustive search used in past works. For wrapper feature selection, we propose the WRANK-ELM, which searches an ordered list of features using the Extreme Learning Machine (ELM) classifier. For hybrid feature selection, we propose the RIG-ELM, which is a filter-wrapper hybrid that uses the Relative Information Gain (RIG) criterion for feature ranking and the ELM for a combinatorial search. Experimental results for different sensor positions and different time resolutions show that the proposed algorithms outperform past work in terms of accuracy and computation time. However, the performance is not so smooth for the complex occupancy profile at a 1-minute time resolution. Dynamic features are more informative than static features, but their use in occupancy estimation has to date been very limited. In this work, we introduce the Feature-Scaled ELM (FS-ELM), which is an ELM-based estimator that uses dynamic features. The FS-ELM is in fact a novel architecture of the ELM, in which a feature layer is added to the standard ELM. The feature layer extracts dynamic features from the raw data. Additionally, we present an effective smoothing strategy to address the problem of noise in environmental parameter data. We demonstrate through experimental results that the FS-ELM yields excellent occupancy estimation accuracy, while retaining the computational efficiency of the standard ELM. Also in this work, we propose a generalized feature selection framework for constructing an occupancy estimator for dynamic features. The framework is a kind of filter-wrapper hybrid feature selection method, which is novel in that it uses a combination of static and dynamic features. In the framework, the filter component works with static features, while the wrapper component works with dynamic features. We use the static features for purposes of speed, since filter methods of feature selection (which work with static features) are quite fast. Thus, the overall computation time of the framework is kept low, while ensuring good accuracy of estimation due to the use of dynamic features. The framework thus offers a reliable method of evaluating a large set of features. To perform occupancy estimation within the proposed framework, we present a novel technique called the Hybrid Feature- Scaled Extreme Learning Machine (HFS-ELM). The HFS-ELM is a dynamic model of the occupancy level in which the present occupancy depends on the measurements of multiple environmental parameters and the estimated occupancy level in a past time horizon. It is thus a generalized version of the FS-ELM, which works with only one type of environmental parameter. Also, while the FS-ELM is limited to two types of arbitrarily selected features, the HFS-ELM accommodates multiple feature types. Through experimental results, we show the excellent performance of the HFS-ELM even for complex occupancy profiles at a 1-minute time resolution.